The Decoder / One term

Can I run AI on my own computer?

Yes: a local AI model is a downloadable file, usually 2 to 10 gigabytes, that runs on your own machine through a free program such as Ollama or LM Studio, with no account, no subscription and nothing sent to a server.

By Mstimaj Numbers checked 30 September 2026 Updated 30 September 2026 Level Technical Tool DeepSeek

Sign in to save your place and mark this done. Free, no card.

A home desktop tower glowing faintly from inside on a dark desk, a single cable unplugged and coiled beside it

You can, today, on an ordinary Windows laptop or desktop with 16 gigabytes of memory. You install one free program, download one model file the size of a couple of films, and type to it in a window. The answers are slower and a little less sharp than the big cloud models, and nothing you type leaves the machine. The one catch: the model has to fit in your memory.

Key facts

  • 16 GBmemory LM Studio recommends for its Windows and Mac apps
  • 2 to 9 GBdownload size of the popular small and medium models on Ollama's library
  • 8 / 16 / 32 GBOllama's original guidance for 7 billion, 13 billion and 33 billion parameter models
  • $0ongoing cost once the file is on your disk
  • LM Studio system requirements page; Ollama library tag pages; Ollama's original README guidance (kept on its docs mirror); checked 30 September 2026

Do I need a graphics card to run AI locally?

No. A graphics card makes it much faster, because the card’s memory and its thousands of small processors are the right shape for this kind of arithmetic, but the small models run on a plain processor. On a CPU only machine expect a reply to arrive at reading pace or a little under, a few words a second, for a model in the 3 to 8 billion parameter range. That is fine for questions, drafts and summaries. It is slow for long documents. LM Studio suggests a card with at least 4 gigabytes of its own memory if you have one; a Mac with Apple silicon counts, because its memory is shared with the graphics side.

What is Ollama, and how does it compare to LM Studio?

They are the two easiest doors in. Ollama is a small background program you drive from the terminal: one command to download a model, one to start chatting, and it also lets other programs on your machine talk to the model, which is how local RAG tools and coding agents plug in. LM Studio is a desktop app with a window, a model browser and a chat pane, and it needs no terminal at all. Both are free. Both use the same model files underneath. If you have never opened a terminal, start with LM Studio. If you want other tools to use the model, Ollama.

Which local model should I download for 16 GB of RAM?

Start with a model in the 3 to 8 billion parameter range, which downloads at 2 to 5 gigabytes and leaves room for everything else you have open. Llama 3.2 3B and Gemma 3 4B are the light ones; Llama 3.1 8B, Qwen 3 8B and Mistral 7B are the sensible middle; DeepSeek R1 8B is the one that shows its reasoning. Gemma 3 12B and Phi-4 14B are worth trying at 16 gigabytes but they will squeeze. Above that you want 32 gigabytes or a graphics card.

Popular local models on Ollama and what they need
Model Download Fits comfortably in Good for
Llama 3.2 3B 2.0 GB 8 GB RAM Quick questions, short drafts
Gemma 3 4B 3.3 GB 8 GB RAM Chat, summaries, some images
Mistral 7B 4.4 GB 16 GB RAM General writing
Llama 3.1 8B 4.9 GB 16 GB RAM General purpose, tool use
Qwen 3 8B 5.2 GB 16 GB RAM Writing, code, many languages
DeepSeek R1 8B 5.2 GB 16 GB RAM Step by step reasoning
Gemma 3 12B 8.1 GB 16 GB, tight Better answers, slower
Phi-4 14B 9.1 GB 16 GB, tight Reasoning and maths
Download size, gigabytes DOWNLOAD SIZE, GIGABYTES Llama 3.2 3B 2.0 GB Gemma 3 4B 3.3 GB Llama 3.1 8B 4.9 GB DeepSeek R1 8B 5.2 GB Gemma 3 12B 8.1 GB Phi-4 14B 9.1 GB
Sizes from Ollama's library pages, 30 September 2026. A film in HD is about 4 GB, so most of these are one or two films.

Is local AI really private?

The model itself sends nothing. It is a file doing arithmetic on your disk, and you can pull the network cable to prove it. What can leak is the program around it: an app that checks for updates, a plugin that searches the web on your behalf, or a chat window that syncs history to a cloud account. Read what the program does on first launch, keep web search off for anything sensitive, and the private claim holds. This is the one setup where a client’s document never touches a vendor.

Can I run ChatGPT offline?

Not ChatGPT itself. That is OpenAI’s service and its models stay on OpenAI’s servers, apart from a family of smaller open weight models the company released for local use. What you can run offline are the open weight models from Meta, Google, Alibaba, Mistral, DeepSeek and others, which are the ones Ollama and LM Studio offer. For everyday questions the 8 billion parameter versions feel like a chat model from a year or two ago. For hard reasoning and long documents, the cloud models are still ahead.

Ask what did not make sense

Comments go through our own form and are read before they appear. Links are not allowed; describe it in words.

Comments are read before they appear.

The newsletter

Get the next term.

One email when something goes up. The argument, the film and the term, nothing else. One click to leave.

Have a question now? Ask Mstimaj, she knows the archive.